CRCNS: Coordinating learning by top-down gating of plasticity in dendrites

在生物学习中有一个中心问题被称为“学分分配问题”：如何 有关决策或行为结果的信息会修改正确神经元中的正确突触 跨越多个大脑区域，以提高未来的表现？中对此问题的标准解决方案 人工神经网络是执行直接梯度下降，这最大限度地减少了输出的误差 通过根据该误差精确地调整每个连接的强度来建立网络。然而，它是 大脑不太可能计算出每个突触对表现误差的影响 “反向传播”细粒度的错误信号，通过多层神经元电路传递到每个突触。 最近的工作为大脑中的监督学习确定了一种新的候选生物学机制。在……里面 除了处理感觉输入的“自下而上”连接外，神经元还发送“自上而下”的信号。 与下层神经元树突的连接。这种反馈推动了特殊活动，称为 诱导一种有效形式的突触可塑性并使神经元成为 只需一次试验，就能选择性地选择刺激功能，这种现象被称为“一次性学习”。这 该项目旨在开发受这些实验观察启发的新学习理论，并 在清醒的、行为正常的小鼠身上实验测试这一理论的预测，以更好地理解自上而下 大脑中的指示信号通过调节多层神经元电路来协调学习 树枝状钙尖峰与相关的可塑性。 该团队结合了神经细胞和突触生理学、系统和计算方面的专业知识。 神经科学和机器学习，以更好地理解一种重要的认知功能-记忆 在目标导向的学习过程中形成。一个主要的目标是开发和批判性地测试一种新的 学习基于树突状钙尖峰和相关突触可塑性的调节。 计算建模将直接为拟议的实验提供信息，这需要成像和 在小鼠的空间觅食行为中在体内操纵神经元群体活动。初步 结果表明，将这些来自生物学的见解融入到人工神经网络中会导致 与标准技术相比，性能更高，突出了 建议的方法。

There is a central problem in biological learning known as the “credit assignment problem”: how does information about the outcome of a decision or behavior modify the right synapses in the right neurons across multiple brain regions to improve future performance? The standard solution to this problem in artificial neural networks is to perform direct gradient descent, which minimizes error in the output of a network by precisely adjusting the strengths of every connection in proportion to that error. However, it is unlikely that the brain is able to compute the impact of each synapse on performance error and “backpropagate” fine-grained error signals across multiple layers of neuronal circuitry to every synapse. Recent work identified a new candidate biological mechanism for supervised learning in the brain. In addition to “bottom-up” connections that process sensory inputs, neurons also send “top-down” connections to the dendrites of neurons in lower layers. This feedback drives special events called “dendritic calcium spikes” that induce a potent form of synaptic plasticity and cause neurons to become selective for stimulus features in as few as a single trial, a phenomenon called “one-shot learning.” This project aims to develop new learning theory inspired by these experimental observations, and to experimentally test predictions of this theory in awake, behaving mice to better understand how top-down instructive signals in the brain coordinate learning across multiple layers of neuronal circuitry by regulating dendritic calcium spiking and associated plasticity. The team synergizes expertise in neuronal cellular and synaptic physiology, systems and computational neuroscience, and machine learning to better understand an important cognitive function - memory formation during goal-directed learning. A major objective is to develop and critically test a new theory of learning based on the regulation of dendritic calcium spikes and associated synaptic plasticity. Computational modeling will directly inform the proposed experiments, which entail imaging and manipulating neuronal population activity in vivo during spatial foraging behavior in mice. Preliminary results suggest that incorporation of these insights from biology into artificial neural networks leads to enhanced performance compared to standard techniques, highlighting the transformative potential of the proposed approach.